Abstract
The LHCb detector is due to be upgraded for processing high-luminosity collisions, which will increase data bandwidth to the event filter farm from 100 GB/s to 4 TB/s, encouraging us to look for new ways of accelerating Online reconstruction. The Coprocessor Manager is a new framework for integrating LHCb's existing computation pipelines with massively parallel algorithms running on GPUs and other accelerators. This paper describes the system and analyzes its performance.
Highlights
Any system for running algorithms within LHCb’s software environment has to adapt to Gaudi
General-purpose GPU computation later yet. It builds on the idea of having each node process events one by one; several Gaudi instances can run concurrently on a single multicore machine as multiple processes, but each instance has its own copies of the transient stores and they do not communicate with each other. This approach to processing events was prudent when the framework was designed, but in the present it stands in the way of making effective use of powerful massively parallel hardware, which is typically heavily optimized for single-instruction multiple-data (SIMD) execution
In order to cooperate with the Coprocessor Manager, GaudiHive needs a way for tasks to suspend themselves
Summary
Other collider experiments have investigated the idea of moving some of their computation to GPUs. Researchers at ALICE (A Large Ion Collider Experiment) at LHC took the step and investigated the issues that arise from GPU algorithm integration with existing software infrastructure [8] They have successfully added GPU track reconstruction to ALICE’s High-Level Trigger (HLT) and the offline reconstruction framework. The server accepts requests and calls whichever kernel is requested in the order it receives the requests This approach is in practice very similar to ALICE’s method of packaging CUDA code into dynamically loaded libraries; in both cases GPU kernel calls are made available to regular Athena algorithms as external procedures. Detectors like LHCb, where individual events require relatively little computation, need massively parallel event processing systems to combine and simultaneously process multiple events This difficulty will increase as long as the growth of GPU performance outpaces growth in information volume produced by particle detectors. This lack of restrictions serves to facilitate porting of existing tested code to the new infrastructure
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
